Photonic Neuron With on Frequency-Domain ReLU Activation Function

Abstract

Driven by an exponential growth of data, neuromorphic computing has risen in popularity as a new method for high-performance computing. The adopted neural network (NN) model relies on parallel processing between neurons and synapses, which reduces the energy consumption and boosts the computational efficiency. Photonics empowers neuromorphic processors through its inherent parallelism, along with high speed and unique bandwidth characteristics. Yet, it requires to transfer each constituent of the NN model to the optical realm, including the challenging nonlinear part of an activation function. Towards this direction, we experimentally demonstrate a photonic rectified linear unit (ReLU) function by employing frequency coding of neural signals in combination with a periodic optical filter. Furthermore, we show that multiple neural sub-circuits can be collapsed over the proposed photonic ReLU hardware and further evaluate the possibility to integrate weighting functionality with the frequency-domain ReLU as a way to further simplify the optical NN. For these demonstrations, we accomplish a low penalty of 1-3% in terms of accuracy when transferring the Iris flower classification challenge from the digital to the optical realm. Finally, we introduce an efficient translucent interface between the linear and nonlinear circuits of a photonic neuron, utilizing an optical frequency-coder that is directly driven by the photocurrent of a preceding photodetector - without the need for electrical amplification.